A systematic review and meta‐analysis of clinical predictors of lithium response in bipolar disorder

Objective To determine clinical predictors of lithium response in bipolar disorder. Methods Systematic review of studies examining clinical predictors of lithium response was conducted. Meta‐analyses were performed when ≥2 studies examined the same potential predictor. Results A total of 71 studies, including over 12 000 patients, identified six predictors of good response: mania‐depression‐interval sequence [odds ratio (OR): 4.27; 95% CI: 2.61, 6.97; P < 0.001], absence of rapid cycling (OR for rapid cycling: 0.30; 95% CI: 0.17, 0.53; P < 0.001), absence of psychotic symptoms (OR for psychotic symptoms: 0.52; 95% CI: 0.34, 0.79; P = 0.002), family history of bipolar disorder (OR: 1.61; 95% CI: 1.03, 2.52; P = 0.036), shorter prelithium illness duration [standardised mean difference (SMD): −0.26; 95% CI: −0.41, −0.12; P < 0.001] and later age of onset (SMD: 0.17; 95% CI: 0.02, 0.36; P = 0.029). Additionally, higher body mass index was associated with poor response in two studies (SMD: −0.61; 95% CI: −0.90, −0.32; P < 0.001). There was weak evidence for number of episodes prior to lithium treatment (SMD: −0.42; 95% CI: −0.84, −0.01; P = 0.046), number of hospitalisations before lithium (SMD: −0.40; 95% CI: −0.81, 0.01; P = 0.055) and family history of lithium response (OR: 10.28; 95% CI: 0.66, 161.26; P = 0.097). Conclusions The relative importance of these clinical characteristics should be interpreted with caution because of potential biases and confounding.


Introduction
Globally, guidelines recommend lithium as first-line maintenance treatment for bipolar disorder (BPD) (1)(2)(3). While lithium has a higher complete response rate than other mood stabiliser medication, only one in three patients will respond well to the drug (4). A number of studies have attempted to identify predictors of response from biological, genetic, clinical and psychosocial characteristics. A recent review of biomarkers to predict lithium response was somewhat discouraging (5). Genome-wide association studies have developed a polygenic risk score for lithium response (6) and large biological marker studies are just beginning (7). However, despite enormous potential to improve our understanding of the lithium-responding subtype of BPD, these approaches are unlikely to be able classify responders accurately without the inclusion of additional clinical features (6). We identified four reviews of multiple clinical lithium response markers, with the most recent attempt to meta-analyses original studies published in 2005 (8)(9)(10)(11). These reviews are limited in their scope as they are not systematic and fail to meet PRISMA standards (12). Factors associated with lithium response described in these reviews include the course of illness, family history of bipolar disorder, family history of lithium response, age at illness onset, number of bipolar hospitalisations, mania-depression-interval (MDI) course sequence, depression-mania-interval (DMI), continuous cycling (CC) (<4 episodes per year without euthymic intervals (13)), rapid cycling (RC) (≥4 episodes per year (13)) and bipolar II disorder (BPD II). We also identified reviews which examined single predictors: pretreatment episode count (14) and episode sequence (15). In the light of these issues, we systematically reviewed the existing literature on clinical predictors of lithium response in BPD and performed meta-analysis where possible.

Methods
This systematic review followed the MOOSE guidelines and PRISMA statement (12,16).

Eligibility criteria
We included randomised trial and observational studies, including adult participants diagnosed with BPD receiving lithium monotherapy. Studies that did not report separate analyses of patients treated with lithium were excluded. Studies examining the use of lithium for other indications (such as unipolar depression) were excluded. We considered studies to be eligible for inclusion if they reported an association between patient level factors (e.g. age at illness onset) and any definition of a lithium response (e.g. recurrence under lithium treatment).

Information sources
We searched EMBASE, Medline and Web of Science from inception to July 2018; the final search was performed on July 14,2018. Additional studies were identified through screening reference lists of included studies and relevant papers. We included only English language studies in humans. Other articles relevant to this topic were searched for via Google Scholar, using reference lists of relevant studies.

Search
We used the following search terms to search all trials registers and databases: [Lithium* OR lithium blood level OR lithium carbonate OR lithium citrate OR treatment response* OR drug response* OR predictor*] AND [Bipolar disorder] AND [observational stud* OR controlled clinical trial* OR RCT OR randomised controlled trial*].

Study selection
Eligibility screening was performed independently by three reviewers. The first author (TPH) screened the titles and abstracts of potential studies to determine inclusion, with a 20% random sample of records independently screened by two reviewers (AK and LS). Eligible studies were subsequently confirmed by the three reviewers (TPH, AK and LS) who independently checked the full text of all retrieved articles. Disagreement was resolved through discussion and consensus between TPH, AK, LS and JFH.

Data collection process
One reviewer (TPH) extracted the following data from included studies and the second (LS) checked the extracted data, including author details, year of publication, types of study design, sample size, interventions investigated, comparison, outcome evaluation or definition of lithium response and key finding. Disagreements were resolved by discussion between TPH, AK, LS and JFH.

Data items
Information was extracted from each included study on: (i) characteristics of study participants (including sample size and number of lithium responders and non-responders; (ii) intervention details (dose, duration of lithium treatment); (iii) definition of a treatment response (number of recurrence under lithium treatment, reduction in time spent in hospital under lithium treatment, reduction of episode frequency, or improvement during lithium treatment based on valid scales, such as Illness severity index (ISI) (17), Affective Morbidity Index (AMI) (18) and ALDA scale (19)); (vi) potential predictors examined; (v) summary results. Data sharing is not applicable to this article as no new data were created or analysed in this study.

Risk of bias in individual studies
Three reviewers (TPH, AK and LS) independently rated each eligible study. The quality of each individual study was evaluated using the modified Downs and Black quality assessment scale (Table S1), which consists of 26 questions to evaluate both randomised and non-randomised studies (20). Question 27 evaluating power was excluded as power should not be part of quality assessment as the aim of a meta-analysis is to detect an effect from inconclusive or underpowered studies. Each criterion is worth one point, and a total score of 20 or above, between 15 and 19, and 14 or below is considered a study of good, fair and poor quality respectively. This quality assessment tool evaluates study reporting, external and internal validity including bias and confounding. Discrepancies between the two reviewers were resolved by discussion and consensus.

Synthesis of results and risk of bias across studies
Meta-analyses were performed after the four assumptions of homogeneity were assessed: (i) studies should be similar in terms of patients recruited; (ii) studies should be comparing the same intervention or exposure with similar controls, (iii) studies should be reporting the same outcomes, (iv) the effect of a predictor should ideally be in the same direction (21). Narrative analysis was carried out along with meta-analysis if only some of the included studies met all of the criteria. For each meta-analysis, where there were two or more studies using the same sample of patients, we excluded the smaller or earlier study.
Meta-analysis using the DerSimonian and Laird random effect model was conducted for each predictor because we assumed heterogeneity existed across different studies, given the definitions of lithium response across studies were inconsistent (22). For binary outcomes, results of the primary studies were summarised as odds ratios (ORs). For continuous outcomes, results of the primary studies were summarised in standardised mean difference (SMD). Pooled ORs or SMDs and corresponding 95% confidence intervals were calculated if two or more studies reported the same clinical predictor. A number of studies categorised patients with BPD as 'partial responder' in addition to 'responder' and 'non-responder'. In order to conduct the random-effects pairwise analysis, we combined the group 'partial responder' and 'non-responder' and formed the group 'partial or non-responder' to avoid chances of data contamination that might impact the results of clinical predictors of lithium responders. Heterogeneity for each predictor was assessed using forest plots and a measure of inconsistency (I 2 ). Publication bias was examined visually through evaluating funnel plots. Stata version 15 was used for all analyses.

Studies included
Our search resulted in 3897 unique citations. Of these, 3670 studies were excluded as the titles and abstracts were not relevant to the research topic, leaving 137 potentially eligible studies for which the full text was reviewed ( Fig. 1). At this stage, 71 studies did not meet the inclusion criteria. An additional five studies that met the inclusion criteria were identified by checking the references of relevant papers and searching via Google Scholar. A total of 71 studies met all inclusion criteria and were included in systematic review, and 44 of these provided data which could be meta-analysed. These studies are described in Table 1. Studies were excluded from the meta-analysis if the population overlapped with another included study population or if it was not possible to calculate the OR or SMD. This meant two large studies using Danish population registers could not be included (23,24).
In total, 19 clinical variables were identified from the articles and further assessed as predictors of lithium response in at least two or more studies: (i) age at study start, (ii) age at illness onset, (iii) prelithium illness duration, (iv) number of episodes prior lithium treatment, (v) number of hospitalisations prior to lithium, (vi) type of BPD (BPD I vs. BPD II), (vii) interval course sequence (MDI vs. DMI), (viii) CC, (ix) irregular sequence (IRR) (absence of any regular mania-depressionsequence), (x) RC, (xi) index episode (mania vs. depression), (xii) predominant polarity (mania vs depression), (xiii) family history of any affective disorder, (xiv) family history of BPD, (xv) family history of lithium response, (xvi) alcohol and drug use, (xvii) psychotic symptoms, (xviii) sex and (xix) body mass index (BMI).

Age at illness onset
A total of 21 studies explored the effect of age at illness onset; five studies (25)(26)(27)(28)(29) were excluded because of insufficient reporting; two studies reported categorical age data rather than continuous data and were therefore not included in meta-analysis. The study by Okuma and colleagues (30) categorised patients into four age groups (>20; 21-30; 31-40; <40) and found no association between age at illness onset and lithium response. However, a similar study conducted by Schurhoff et al. (31) found that late onset (40 years old or older) was associated with good lithium response (P = 0.04). Pooling the remaining 14 eligible studies, with a total sample of 2063 patients, there was an association between age at onset and treatment response (SMD = 0.17; 95% CI: 0.02 to 0.33; P = 0.029; Fig. 2, Figure S1), but heterogeneity was high (I 2 = 58.3.6%; P = 0.003). Of these included studies, four found increasing age was associated with increased chance of lithium response (32)(33)(34)(35) and one found increased age was associated with a reduced chance of response (4).

Age at study start
The association between age at study start and lithium treatment response was quantified in 10 studies with a total sample of 1266 patients. A medium level heterogeneity was observed (I 2 = 50.8%; P = 0.032). The pooled effect estimate suggested no association between study admission age and lithium response (SMD: 0.02; 95% CI: À0.17 to 0.21; P = 0.851; Fig. 2, Figure S1).

Prelithium illness duration
Data from five studies with a sample of 931 patients were pooled (Table 1). Heterogeneity was low (I 2 = 0.0%; P = 0.701). The results suggested that a short prelithium treatment illness duration was associated with good lithium response (SMD = À0.26; 95% CI: À0.41 to À0.12; P < 0.001; Fig. 2, Figure S2). This was also true in the study by Kessing and colleagues of 4714 individuals with BPD (24); those commenced on lithium at first contact had lower rates of non-response compared to those commenced at later contacts (HR 0.87, 95% CI 0.76 to 0.91, P < 0.0001).  Patients whose families showed high expressed emotion were over-represented in the poorer outcome groups (P = 0.004) as were those whose families had a negative affective style (P Years of education Alda treatment response scale Responders had fewer episodes prior to lithium (P = 0.012). sex (P = 0.379), age at study start (P = 0.993), education (P = 0.876), age at onset (P = 0.837), family history of BPD (P = 0.708) and psychotic features (P = 0.698) were not associated with response FAIR BPD, bipolar disorder; CC, continuous cycling; CI, confidence interval; DMI, depression-mania-interval sequence; HR, hazard ratio; IRR, irregular sequence; MDI, mania-depression-interval sequence; na, not available; NS, non-significant (at P = 0.05); PTSD, posttraumatic stress disorder; RC, rapid cycling; RR, risk ratio.*Data included in meta-analysis.

Number of episodes prior lithium treatment
The impact of mean number of episodes prior to lithium treatment on treatment response was assessed in seven studies with a total sample of 824 (Table 1). Meta-analysis suggested that increased number of mood episodes prior to commencing lithium was weakly associated with reduced chance of good response (SMD = À0.42; 95% CI: À0.84 to À0.01; P = 0.046; Fig. 2, Figure S3). Heterogeneity was high (I 2 = 85.9%; P < 0.001).

Number of hospitalisations prior to lithium treatment
A combined sample of 673 patients from four studies contributed data on number of previous hospitalisations. Although two studies suggested fewer hospitalisations were associated with good response (36,37), overall there was no evidence of a clear association between number of hospitalisations and lithium response SMD = À0.40; 95% CI: À0.81 to 0.01; P = 0.055; Fig. 2, Figure S3). In the Danish population (23), increasing number of hospitalisations between diagnosis and starting lithium were associated with increased rates on non-response (HR 1.03, 95% CI 1.02 to 1.05, P = 0.0002).

Type of bipolar disorder
The association between BPD subtype and good lithium response was quantified in 11 studies with a total of 1556 patients. There was evidence of considerable heterogeneity (I 2 = 70.7%; P < 0.001) across studies, and the result indicated insufficient evidence to support BPD I as a clinical predictor of lithium response when comparing to patients with BPD II (OR: 1.01; 95% CI: 0.58 to 1.76; P = 0.971; Fig. 3, Figure S4). At an individual level, two of the included studies suggested BPD I may be associated with a preferential lithium response (32,34) and three suggested BPD II may be associated with a preferential lithium response (4,33,38).

Continuous cycling
The impact of continuous cycling on lithium treatment response was quantified in seven studies with a total of 804 patients. Meta-analysis suggested no association between continuous cycling and response (OR: 0.65; 95% CI: 0.34 to 1.26; P = 0.204; Fig. 3, Figure S6).

Rapid cycling
The impact of the presence of RC on lithium treatment response was quantified in nine studies with a total of 1442 patients. Moderate heterogeneity was identified (I 2 = 37.5.6%; P = 0.119). The meta-analysis result indicated evidence that patients displaying RC have reduced odds of lithium response compared to those without RC (OR: 0.30; 95% CI: 0.17 to 0.53; P < 0.001; Fig. 3, Figure S6).

Polarity of index episode
There was no evidence of an association between lithium response and manic index episode (OR: 1.12; 95% CI: 0.56 to 2.21; P = 0.753; Fig. 3, Figure S7). From six studies, one suggested a manic index episode was a good predictor of response (32) and one suggested a depressive index episode was a good predictor (33). Others were inconclusive, and heterogeneity was high (I 2 = 73.7%; P = 0.002). Kessing et al. found reduced rates of non-response in individuals with a manic index episode (HR 0.84, 95% CI 0.77 to 0.91) and elevated rates in those with a depressive index episode (HR 1.13, 95% CI 1.03 to 1.25) compared to those whose index episode was 'remission, other or unspecified'. However, it is unclear who is included in this reference category and there is potential misclassification because of the routine registerbased nature of the data source.

Predominant mood polarity
Predominant mania or depression was documented in three studies with a total sample of 280 patients. Overall, there was no evidence for an association between lithium response and mania over depression dominance (OR: 1.07; 95% CI: 0.07 to 15.74; P = 0.959; Fig. 3, Figure S8).

Family history
Eight studies, including 714 individuals, contributed to meta-analysis of the association between family history of any affective disorder and lithium response. There was no evidence of an association (OR: 1.13; 95% CI: 0.75 to 1.69; P = 0.560; Fig. 3, Figure S9). Individuals with a family history of bipolar disorder were more likely to have a good response to lithium (10 studies, 1454 patients; OR: 1.61; 95% CI: 1.03 to 2.52; P = 0.036; I 2 = 43.5%; heterogeneity P = 0.068; Fig. 3, Figure S9). One study, which could not be combined in meta-analysis, runs contrary to this, finding 88% of individuals without a family history have a reduction in episode frequency during lithium treatment, while only 68% of those with a family history of BPD. Only two studies (79 patients) could be included in meta-analysis of family history of lithium response. Both studies had point estimates suggesting good lithium response in family members may be associated with good response in the index patient, however, confidence intervals overlapped no effect (OR: 10.28; 95% CI: 0.66 to 161.26; P = 0.097, Fig. 3, Figure S9).

Alcohol and drug use
The association between alcohol and drug use and lithium response was investigated in three studies with a total sample of 540 patients. The results showed a medium heterogeneity (I 2 = 54.5%; P = 0.111) and demonstrated no evidence to suggest alcohol and drug use as a potential predictor of lithium response (OR: 0.55; 95% CI: 0.23 to 1.34; P = 0.189; Fig. 3, Figure S10).

Psychotic symptoms
A total sample of 1066 patients from eight studies were included in assessing psychotic symptoms. Medium heterogeneity was observed (I 2 = 42.8%; P = 0.093), and the result suggested a strong association between psychotic symptoms and poor response (OR: 0.52; 95% CI: 0.34 to 0.79; P = 0.002; Fig. 3, Figure S11).

Sex
The role of sex as a potential lithium response predictor was investigated 1,729 patients from 17 studies. Sex was not associated with lithium treatment response (being male OR: 0.89; 95% CI: 0.68 to 1.15; P = 0.356; I 2 = 22.7%; heterogeneity P = 0.191; Fig. 3, Figure S12). However, the only population-based study identified suggested an association between being female and non-response (HR 1.12, 95% CI 1.04 to 1.21, P = 0.002) (23).

Further potential predictors
A study by Rybakowski et al. investigated the relationship between temperament and lithium response (39). Data from 71 patients suggested that lithium response was correlated positively with hyperthymic score (r = 0.31; P = 0.009), and negatively with anxiety and cyclothymic temperament scores (r = À0.27; P = 0.022 and r = À0.26; P = 0.032 respectively). We identified one other study which examined personality traits and treatment response (40). This study reported that responders had higher dominance scores (Pvalue < 0.05), lower neuroticism scores (Pvalue < 0.01) and were less likely to have 'deviant personalities' (P-value < 0.05). Social support was examined in two studies with overlapping study populations (36,41) and a third study which presented results in a way that did not permit metaanalysis. Lower social support was associated with poor response in each case. Other sociodemographic characteristics were reported in a small number of studies. Social class was associated with response in one identified study, but not in another (36,42). Education, marital status, (38) and ethnicity (43) were not associated with lithium response. Employment status was associated with response in one large nationwide population study (23), but not in a smaller observational study (38). Insulin resistance was found to be associated with poor response to lithium in one study, in keeping with the studies showing an association with BMI (47)(48)(49). While we did not consider childhood trauma as a 'clinical' predictor of treatment response, one included study examined this among other features (50). This study suggested physical abuse was an independent predictor of poor lithium response after accounting for many clinical characteristics. However, the only other study we could identify examining childhood trauma found no association between lithium response and any type of trauma (51).

Risk of bias within studies
Overall, the mean Downs and Black quality assessment score was 16.3, which is considered fair quality. We identified eight good quality studies, 45 fair quality studies and 18 poor quality studies (Tables S1 and S2). Most of the studies failed to report or account for appropriate confounders in regression analyses.

Risk of bias across studies
In line with the Sterne et al. (52), funnel plot asymmetry was assessed when 10 or more studies were included in the meta-analysis. Funnel plots were produced for age at illness onset ( Figure S14), sex ( Figure S15), family history of BPD ( Figure S16), age at study start ( Figure S17) and type of BPD ( Figure S18). The studies of BPD subtype, sex and family history produced asymmetrical funnel plots. A possible source of this asymmetry is true heterogeneity between studies; potentially because of differences in lithium dosage, treatment duration or diagnostic definition, small sample sizes and the low number of studies included.

Discussion
We identified a total of 71 studies, including over 12 000 patients which explore clinical predictors of lithium treatment response in patients with BPD. From these, six predictors of good response were identified. Our results suggest that predictors of good response are (i) MDI sequence, (ii) absence of RC, (iii) absence of psychotic symptoms, (iv) shorter prelithium illness duration, (v) family history of bipolar disorder and (vi) later illness onset. Additional features which may be related to response are body mass index, number of episodes before lithium treatment, number of hospitalisations before lithium and family history of lithium response.
Our findings generally correspond with previous review articles (8)(9)(10)(11). As far as we are aware, Kleindienst et al. conducted the only previous meta-analysis of multiple clinical response predictors and our results were broadly similar (8).
However, we did not find a strong association with number of previous hospitalisations or CC, and they found no association with prelithium illness duration, psychotic symptoms or RC. This may be because of differing approaches to study inclusion and analysis, and in some cases because contradictory results have been found in individual studies published since 2005. Additionally, prelithium illness duration, number of episodes prior to lithium treatment and number of hosptialisations prior to lithium are likely to all be measuring a similar underlying concept.
Clinically, these predictors are likely to be of varying importance. Some may essentially reflect establishing a more benign illness course because of early intervention and may not be specific to lithium. This may be the case for shorter prelithium illness duration, and fewer episodes prior to lithium treatment, which are clearly related to illness severity. Others may be more central to guiding the choice to use lithium. DMI sequence, rapid cycling and psychotic symptoms are all associated with poor lithium response, so their presence may suggest an alternative treatment might be more appropriate for the patient. However, there is limited evidence to suggest any other drug therapy would lead to better than responses than with lithium. Family history of bipolar disorder and potentially family history of lithium response (likely under powered in our analysis) are important as they may reflect a more heritable subtype of BPD.

Limitations
The reliability of the potential predictors identified remains unclear. For most of the meta-analyses conducted, estimates were highly heterogeneous, often including studies suggesting both a positive and negative effect of the predictor. Most studies were rated as fair or poor in terms of quality. Often insufficient statistical information was reported in the primary study to conduct meta-analysis; most studies failed to report adequate summary statistics such as standard deviation or number of responders and non-responders. Sample sizes were often small and studies consisted of highly selective groups of patients. Also, the definition of lithium response in many of the studies did not rely on a standardised tool, which can greatly influence the process of identifying lithium responders and lithium non-responders. As shown in Table 1, most of the studies relied on recurrence of an affective episode under lithium treatment to define lithium non-responders. However, this definition of lithium response fails to consider changes in episode frequency or symptom severity, and so may miscategorise responders and non-responders. Scott and colleagues note that using continuous scores for lithium response as opposed to categories of response leads to different predictors being identified (53). Additionally, none of the studies reported lithium plasma level or adherence to treatment by response status. Information on these factors would strengthen the argument that these are true predictors of lithium response as it would then be possible to rule out differences in the way treatment is used as a cause of the observed associations.
Very few of the studies explored the possibility of interdependence or interaction between predictors. For example, interdependence might exist between prelithium illness duration and illness severity (54). A greater illness severity is related to receiving early treatment and subsequently decreasing illness morbidity. Accordingly, a short prelithium illness duration might appear to be related to good lithium response (54). Only some of the more recent studies included multiple covariates in the same model (for example; (23,50,53)) an approach which is necessary to determine whether covariates are truly independent predictors.
Because of the low reliability of the results and the inability to eliminate biases, any clinical conclusions relating to any single predictor should be made cautiously. Because of the limitations of the data, particularly the limited number of RCTs, it is difficult to separate predictors of lithium response from predictors of a benign illness course.
In conclude although we identified six potential clinical predictors of lithium response, there are a number of issues relating to their reliability and validity which cannot be addressed by reviewing the existing literature. As with response classification by genetic or biological markers, clinical response prediction is likely to be complex and multivariable. Studies need to explore multiple predictors, and their interactions, with operationalised end points for lithium response.  Table S1. Modified Downs and Black checklist for the assessment of methodological quality of both randomized and nonrandomized studies. Table S2. Study quality scores using modified Downs and Black scale: Checklist for measuring study quality (n = 50). Figure S1. Relationship between age at illness onset, age at study start and lithium treatment response. Figure S2. Relationship between pre-lithium illness duration and lithium treatment response. Figure S3. Relationship between number of episodes, number of hospitalizations prior lithium treatment and lithium treatment response. Figure S4. Relationship between bipolar I disorder, bipolar II disorder and lithium treatment response. Figure S5. Relationship between types of episodic sequence and lithium treatment response. Figure S6. Relationship between types of cycling and lithium treatment response. Figure S7. Relationship between depressive index episode, manic index episode and lithium treatment response. Figure S8. Relationship between predominant mood polarity and lithium treatment response. Figure S9. Relationship between types of family history and lithium treatment response. Figure S10. Relationship between alcohol and drug use and lithium treatment response. Figure S11. Relationship between psychotic symptoms and lithium treatment response. Figure S12. Relationship between sex and lithium treatment response. Figure S13. Relationship between Body mass index and lithium treatment response. Figure S14. Funnel plot of studies examining age at illness onset as a predictor of lithium response, with pseudo 95% confidence limits. Figure S15. Funnel plot of studies examining sex as a predictor of lithium response, with pseudo 95% confidence limits. Figure S16. Funnel plot of studies examining family history of bipolar disorder as a predictor of lithium response, with pseudo 95% confidence limits. Figure S17. Funnel plot of studies examining age at study start as a predictor of lithium response, with pseudo 95% confidence limits. Figure S18. Funnel plot of studies examining bipolar disorder subtype as a predictor of lithium response, with pseudo 95% confidence limits.